It seems that training dataset is full of lightly to hard mislabelled masks. Most of them are with missed fins and antennas on the top of cars or brutal rectangles at the bottom instead of accurate segmentations. Sometimes it looks like sabotage from the staff that don't want to be replaced by AI or their cats jumping on the keyboard like the one attached (have your ever checked all the masks?).
\n\nI have checked first 1000 images and the labels that look bad for me are listed in broken.txt.
\n\nLet's find more images meant to be fixed together!
", "messages": [ { "id": "213264", "postDate": "08/14/2017 09:23:55", "content": "It seems that training dataset is full of lightly to hard mislabelled masks. Most of them are with missed fins and antennas on the top of cars or brutal rectangles at the bottom instead of accurate segmentations. Sometimes it looks like sabotage from the staff that don't want to be replaced by AI or their cats jumping on the keyboard like the one attached (have your ever checked all the masks?).
\n\nI have checked first 1000 images and the labels that look bad for me are listed in broken.txt.
\n\nLet's find more images meant to be fixed together!
", "rawMarkdown": "It seems that training dataset is full of lightly to hard mislabelled masks. Most of them are with missed fins and antennas on the top of cars or brutal rectangles at the bottom instead of accurate segmentations. Sometimes it looks like sabotage from the staff that don't want to be replaced by AI or their cats jumping on the keyboard like the one attached (have your ever checked all the masks?).\n\nI have checked first 1000 images and the labels that look bad for me are listed in broken.txt.\n\nLet's find more images meant to be fixed together!", "votes": null }, { "id": "213266", "postDate": "08/14/2017 09:30:33", "content": "No obstacles will stop AI from whatever it wants to achieve
", "rawMarkdown": "No obstacles will stop AI from whatever it wants to achieve", "votes": null }, { "id": "213269", "postDate": "08/14/2017 09:35:16", "content": "Here's my bit, but there is a discussion on that subject already:\nhttps://www.kaggle.com/c/carvana-image-masking-challenge/discussion/37229
", "rawMarkdown": "Here's my bit, but there is a discussion on that subject already:\nhttps://www.kaggle.com/c/carvana-image-masking-challenge/discussion/37229", "votes": null }, { "id": "213273", "postDate": "08/14/2017 09:40:46", "content": "Ok, this one is fine too, the title is what misled me.
", "rawMarkdown": "Ok, this one is fine too, the title is what misled me.", "votes": null }, { "id": "213277", "postDate": "08/14/2017 09:50:03", "content": "Unfortunately there are mistakes and inconsistencies in the train set, eg. the wheel spokes are sometimes segmented and others they're not even on the same car at different angles. The same mistakes are likely present in the test set so I believe the top entries will compete on how well they can predict human error rather than the optimal masks.
", "rawMarkdown": "Unfortunately there are mistakes and inconsistencies in the train set, eg. the wheel spokes are sometimes segmented and others they're not even on the same car at different angles. The same mistakes are likely present in the test set so I believe the top entries will compete on how well they can predict human error rather than the optimal masks.", "votes": null }, { "id": "213295", "postDate": "08/14/2017 10:58:02", "content": "On the other hand - this competition has been running for how many? 3 weeks already?\nA bunch of people spent quite a sum of money for cloud services etc and it would not be nice to tell them to start from a scratch again
", "rawMarkdown": "On the other hand - this competition has been running for how many? 3 weeks already?\nA bunch of people spent quite a sum of money for cloud services etc and it would not be nice to tell them to start from a scratch again", "votes": null }, { "id": "213373", "postDate": "08/14/2017 16:14:53", "content": "1e89e1af42e7_07_mask.gif\n2a4a8964ebf3_08_mask.gif\n2cb06c1f5bb1_04_mask.gif
", "rawMarkdown": "1e89e1af42e7_07_mask.gif\n2a4a8964ebf3_08_mask.gif\n2cb06c1f5bb1_04_mask.gif", "votes": null } ], "comments": [ { "id": 213266, "author_name": "venheads", "author_url": "", "post_date": "08/14/2017 09:30:33", "content": "No obstacles will stop AI from whatever it wants to achieve
", "votes": null, "replies": [] }, { "id": 213269, "author_name": "killthekitten", "author_url": "", "post_date": "08/14/2017 09:35:16", "content": "Here's my bit, but there is a discussion on that subject already:\nhttps://www.kaggle.com/c/carvana-image-masking-challenge/discussion/37229
", "votes": null, "replies": [ { "id": 213273, "author_name": "nizhibitsky", "author_url": "", "post_date": "08/14/2017 09:40:46", "content": "Ok, this one is fine too, the title is what misled me.
", "votes": null, "replies": [] } ] }, { "id": 213277, "author_name": "petrosgk", "author_url": "", "post_date": "08/14/2017 09:50:03", "content": "Unfortunately there are mistakes and inconsistencies in the train set, eg. the wheel spokes are sometimes segmented and others they're not even on the same car at different angles. The same mistakes are likely present in the test set so I believe the top entries will compete on how well they can predict human error rather than the optimal masks.
", "votes": null, "replies": [] }, { "id": 213295, "author_name": "venheads", "author_url": "", "post_date": "08/14/2017 10:58:02", "content": "On the other hand - this competition has been running for how many? 3 weeks already?\nA bunch of people spent quite a sum of money for cloud services etc and it would not be nice to tell them to start from a scratch again
", "votes": null, "replies": [] }, { "id": 213373, "author_name": "svesal", "author_url": "", "post_date": "08/14/2017 16:14:53", "content": "1e89e1af42e7_07_mask.gif\n2a4a8964ebf3_08_mask.gif\n2cb06c1f5bb1_04_mask.gif
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "213264": "It seems that training dataset is full of lightly to hard mislabelled masks. Most of them are with missed fins and antennas on the top of cars or brutal rectangles at the bottom instead of accurate segmentations. Sometimes it looks like sabotage from the staff that don't want to be replaced by AI or their cats jumping on the keyboard like the one attached (have your ever checked all the masks?).\n\nI have checked first 1000 images and the labels that look bad for me are listed in broken.txt.\n\nLet's find more images meant to be fixed together!", "213266": "No obstacles will stop AI from whatever it wants to achieve", "213269": "Here's my bit, but there is a discussion on that subject already:\nhttps://www.kaggle.com/c/carvana-image-masking-challenge/discussion/37229", "213273": "Ok, this one is fine too, the title is what misled me.", "213277": "Unfortunately there are mistakes and inconsistencies in the train set, eg. the wheel spokes are sometimes segmented and others they're not even on the same car at different angles. The same mistakes are likely present in the test set so I believe the top entries will compete on how well they can predict human error rather than the optimal masks.", "213295": "On the other hand - this competition has been running for how many? 3 weeks already?\nA bunch of people spent quite a sum of money for cloud services etc and it would not be nice to tell them to start from a scratch again", "213373": "1e89e1af42e7_07_mask.gif\n2a4a8964ebf3_08_mask.gif\n2cb06c1f5bb1_04_mask.gif" }, "source": "meta" }